OpenAI researcher on why soft skills are the future of work | Karina Nguyen

Karina Nguyen, OpenAI researcher who built Canvas and Tasks, argues soft skills and synthetic data are the real frontiers of AI product development.

• Nguyen left engineering for research after realizing Claude was getting good enough at frontend coding to replace her own work.
• Synthetic data — model-generated training examples — was the primary method used to build Canvas and Tasks, not human-labeled data.
• The data wall narrative is wrong: post-training via reinforcement learning on infinite tasks has no scaling ceiling, only evaluation bottlenecks.
• Canvas was built around three synthetically trained behaviors: when to trigger, how to make targeted edits, and how to add document comments.
• OpenAI is more bottoms-up and risk-taking; Anthropic is more focused on craft, prioritization, and model personality — Claude’s librarian character reflects its creators.
• Nguyen argues prioritization, creativity, and people management are the hardest skills to teach models and will be most durable for humans.
• Computer-use agents are hard because models operate on pixels rather than language, and correctly inferring human intent before acting remains unsolved.
• PhD-level benchmarks like GPQA are saturating above 60–70%, meaning evaluation quality — not training data — is now the bottleneck for frontier models.

2025-02-09 · Watch on YouTube